Large Language Models (LLMs) have been identified as having the unrivaled potential to support educators and learners in generating high-quality educational content at various levels of education. Despite their remarkable capabilities, LLMs’ concerns revolve around the lack of specific domain expertise, their propensity to generate plausible-looking content that is full of error manifested in their hallucinatory behaviors, their tendency to perpetuate bias, and their lack of knowledge temporality, which can profoundly impact the learning experiences of learners in fragile contexts such as refugee communities. This work explores leveraging Retrieval-Augmented Language Models (REALMs) in creating culturally sensitive learning content for refugee learners. Using REALMs for content generation offers an innovative and cost-effective solution that overcomes the core LLM challenges, including hallucinations, lack of domain specificity, bias, and recency issues, while addressing the unique educational challenges of marginalized communities. We discuss the applications of LLMs in creating personalized learning content, assessment item generation, automated lesson plans, schemes of work, syllabus generation, and learning content language translation. Using evidence from literature, we offer transferable use cases to refugee education. We also provide the limitations of LLMs in this task and the associated ethical concerns. Furthermore, we illuminate the processing logic of REALMs and how they overcome some limitations. Ultimately, we provide crucial insights for educators, policymakers, and AI developers on leveraging the transformative capabilities of REALMs to close the learning gaps for refugee learners and promote inclusive education, a critical aspect of Society 5.0.